Grounding, escalation
and a recorded response.
Preview access and known failures
A service call reflects its chosen configuration and can fail. Model agreement is not an independent proof of truth. Keep request/response and billing contracts separate from scientific validity.
Photon combines a locally served model, registry checks and escalation paths. A response can include a signed record of what the system returned. The preview has known coverage and wording failures; agreement between models is not an independent proof of truth.
a live demo you can run yourself — or your own private stack, on your hardware, your data never leaves.
A composed system, with separate evidence for each component.
Registry checks and cross-model agreement can contribute to a response, with observed coverage and wording failures. The activation pre-filter is a routing signal. Inspect the response and route trace under the selected configuration.
the question sets the depth — try every rung in the live demo below.
Ground a claim, live.
This Photon preview checks submitted claims against supported registries and can return a signed receipt. Review the service’s actual access and billing contract before use. The public evidence page preserves observed phrasing failures.
python3 aperture_verify.py verify receipt.json checks any receipt or cert offline against the pinned registry key, with no call back to us. You can't watch a 12-hour agent, but you can verify its trail: verify-session checks every receipt in a run and the root over their ordered sequence, so altering, reordering, or dropping one is caught. Where a fixed rule settles the answer, a deterministic per-answer witness lets you re-derive the verdict yourself — exact, offline, still provenance, not truth. Specs: SPEC · WITNESS_SPEC.The demo runs on ours. Run the real thing on yours.
A Photon deployment requires configuration, model access, registry coverage and data-flow review. Signed records support provenance; they do not establish safe or correct answers.
- the glass-box lens — reads activations, not just tokens
- signed, offline-checkable receipts + a transparency log
- deterministic per-answer witnesses — code · math · registry
- your model, calibrated + its own certificate
- air-gapped — your keys, your hardware, your data
- the full cascade: ground · pair · escalate · abstain
Provenance, not truth — it proves what was served and lets a third party re-check it, on hardware you control.
One key. The whole cascade, in your stack.
The same Photon — grounding and inference — through an OpenAI-compatible API, an SDK, or an MCP tool your agent calls before it commits. One sk-apt- key, one base_url. build on the lens →
Photon, on your hardware.
The whole honesty layer, self-hosted — the glass-box lens, signed receipts anyone re-checks offline, and a null-validated calibration certificate for your model. Your data never leaves your VPC.
- the glass-box lens, verifiable receipts + a transparency log, per-answer witnesses
- your model calibrated first-party, with its own certificate
- air-gapped deploy — your keys, your hardware
Provenance, not truth — a receipt proves what was served and lets a third party re-check it; it never certifies the answer is true.